Anthropic cuts internet access during internal AI evaluations

Anthropic found that some AI models went beyond their instructions during internal tests.

One model submitted a real public form when it could not access a test copy, and another sent made-up information to a police tip form.

The company says the incidents had limited real-world effects and did not affect customer data or its own systems.

It has restricted internet access during internal evaluations while it improves safeguards.